Add Bulgarian Reasoning Dataset (#3137)

Part of this: #3122
https://huggingface.co/datasets/0x22almostEvil/reasoning_bg_oa
---
# Dataset Card for Bulgarian QnA reasoning with ~2.7K entries.

### Dataset Summary

Contains Parquet of a list of instructions and answer.

Each row consists of

* INSTRUCTION
* RESPONSE
* SOURCE (reasoning_bg)
* METADATA (json with language, url, id).

### Original Dataset is available here:
* https://huggingface.co/datasets/reasoning_bg

---------

Co-authored-by: 0x22almostEvil <0x22almostEvil>

data/datasets/__init__.py

@@ -29,6 +29,7 @@
     "ru_riddles_337": "0x22almostEvil/ru-riddles-377",
     "instructional_codesearchnet_python": "Nan-Do/instructional_codesearchnet_python",
     "tatoeba_mt_qna_oa": "0x22almostEvil/tatoeba-mt-qna-oa",
+    "reasoning_bg_oa": "0x22almostEvil/reasoning_bg_oa",
 }
 
 SAFETY_DATASETS = {

data/datasets/reasoning_bg_oa/README.MD

@@ -0,0 +1,21 @@
+# Dataset Card for Bulgarian QnA reasoning with ~2.7K entries.
+
+### Dataset Summary
+
+- License: apache-2.0
+- Contains Parquet of a list of instructions and answers in Bulgarian language.
+
+Each row consists of
+
+- INSTRUCTION
+- RESPONSE
+- SOURCE (reasoning_bg)
+- METADATA (json with language, url, id).
+
+### Link:
+
+- https://huggingface.co/datasets/0x22almostEvil/reasoning_bg_oa
+
+### Original Dataset is available here:
+
+- https://huggingface.co/datasets/reasoning_bg

data/datasets/reasoning_bg_oa/data_process.py

@@ -0,0 +1,48 @@
+import json
+from dataclasses import dataclass
+
+import pandas as pd
+from datasets import concatenate_datasets, load_dataset
+
+configs = ["biology-12th", "philosophy-12th", "geography-12th", "history-12th", "history-quiz"]
+datasets = []
+
+
+@dataclass
+class QnA:
+    INSTRUCTION: str
+    RESPONSE: str
+    SOURCE: str
+    METADATA: str
+
+
+# format in QnA
+def create_qna(row):
+    instruction = f'{row["question"]} {", ".join(row["answers"])}?'
+    response = row["correct"].translate(str.maketrans("", "", "();"))
+    source = "reasoning_bg"
+    metadata = {
+        "language": "bg",
+        "url": f'{row["url"]}',
+        "id": f'{row["id"]}',
+    }
+    metadata_str = json.dumps(metadata)
+    return QnA(instruction, response, source, metadata_str)
+
+
+# merge dataset configs into one
+for config in configs:
+    dataset = load_dataset("reasoning_bg", config, split="train")
+    datasets.append(dataset)
+
+merged_dataset = concatenate_datasets(datasets)
+
+print(merged_dataset)
+
+# convert the dataset to a pandas dataframe
+df = pd.DataFrame(merged_dataset)
+
+qna_list = df.apply(create_qna, axis=1).tolist()
+
+qna_df = pd.DataFrame(qna_list, columns=["INSTRUCTION", "RESPONSE", "SOURCE", "METADATA"])
+qna_df.to_parquet("reasoning-bg-oa.parquet", row_group_size=100, engine="pyarrow", index=False)